DDF2Pol: A Dual-Domain Feature Fusion Network for PolSAR Image Classification

2026-04-22 · Source: cs.CV updates on arXiv.org · Field: Science & Research — Environmental Science & Earth Systems, Engineering & Applied Sciences, Artificial Intelligence & Machine Learning · Depth: Expert, long

Summary

DDF2Pol is a novel, lightweight dual-domain convolutional neural network designed for PolSAR image classification. This architecture features two parallel streams: one processes real-valued feature descriptors, and the other handles complex-valued PolSAR data, both utilizing 3D convolutional layers to capture spatial and polarimetric information. The model further refines features using a depth-wise convolution layer for spatial enhancement and a coordinate attention mechanism to focus on informative regions. Evaluated on the Flevoland and San Francisco benchmark datasets, DDF2Pol achieved an Overall Accuracy (OA) of 98.16% and 96.12% respectively, outperforming several existing real- and complex-valued models. With only 91,371 parameters, DDF2Pol offers an efficient solution, demonstrating strong performance even with limited training data, such as 0.25% of labeled samples.

Key takeaway

For Computer Vision Engineers developing PolSAR classification systems, DDF2Pol demonstrates that combining real and complex-valued feature streams with efficient spatial refinement and attention mechanisms can yield superior accuracy with significantly reduced model complexity. You should consider adopting a dual-domain approach and incorporating depth-wise convolutions and coordinate attention, especially when working with limited labeled data, to achieve robust and scalable classification performance.

Key insights

DDF2Pol fuses real and complex PolSAR features with attention for efficient, high-accuracy classification.

Principles

• Dual-domain processing captures complementary PolSAR features.
• Depth-wise convolution refines spatial features efficiently.
• Coordinate attention enhances feature discrimination.

Method

DDF2Pol employs parallel 3D CNN streams for real and complex PolSAR data, followed by depth-wise convolution for spatial refinement and a coordinate attention block to emphasize relevant features before global average pooling for classification.

In practice

• Use 3D convolutions for efficient PolSAR feature extraction.
• Integrate depth-wise convolution to reduce parameter count.
• Apply coordinate attention for improved accuracy with low overhead.

Topics

• PolSAR Image Classification
• Dual-Domain CNN
• Complex-Valued CNNs
• Depthwise Convolution
• Coordinate Attention

Code references

• mqalkhatib/DDF2Pol

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Frequency-Enhanced Dual-Subspace Networks for Few-Shot Fine-Grained Image Classification — FEDSNet improves few-shot fine-grained image classification by integrating frequency domain features to enhance structural stability and reduce overfitting.
• Polyp-D2ATL: Deep Domain-Adaptive Transfer Learning for Colorectal Polyp Classification under Label Distribution Shift — Polyp-D2ATL uses deep domain-adaptive transfer learning to classify colorectal polyps, effectively mitigating label distribution shift and data imbalance.
• Improved Knowledge Distillation for Land-Use Image Classification — Combining hard and soft supervision with KL divergence and Cosine Similarity improves knowledge distillation for land-use classification.
• Architecture-Adaptive Uncertainty Fusion for Deepfake Detection — Correlation-Optimized Fusion (COF) adaptively combines multiple uncertainty sources to predict deepfake detection errors, improving cross-domain reliability.
• Entropy-Guided Self-Supervised Learning for Medical Image Classification — Combining ImageNet and entropy-guided MAE pre-training via ensemble significantly boosts medical image classification accuracy and robustness.
• DB-FGA-Net: Dual Backbone Frequency Gated Attention Network for Multi-Class Brain Tumor Classification with Grad-CAM Interpretability — DB-FGA-Net offers robust, interpretable, and augmentation-free brain tumor classification using dual-backbone and frequency-gated attention.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.